Enhancing brain tumor classification with a diffusion denoising model and a conditional deep convolutional neural network

Abstract

The limited availability of medical imaging datasets and concerns over patient privacy pose significant challenges in artificial intelligence-driven disease diagnosis. To overcome these limitations, this study introduces the use of the denoising diffusion model (DDM) for generating synthetic datasets, marking a significant advancement over traditional generative adversarial networks (GANs). This research pioneers the integration of DDM with conditional deep convolutional neural networks (CDCNN) for brain tumor classification, focusing on four categories: Glioma, meningioma, pituitary tumors, and healthy tissue. The proposed CDCNN model, developed from existing convolutional neural network architectures, effectively processed both DDM-generated synthetic datasets and original datasets sourced from the Kaggle repository. The results demonstrate the remarkable efficacy of the DDM-based augmentation framework, with the CDCNN model achieving an accuracy of 96.2%, significantly outperforming traditional GAN-based models, such as Pix2Pix. A comparative analysis against established architectures, including ResNet50, Visual Geometry Group (VGG)16, VGG19, and InceptionV3, further highlights the superior sensitivity, specificity, and F1 score of the proposed framework. These findings underscore the transformative potential of diffusion models in enhancing dataset diversity, improving classification performance, and addressing data scarcity issues in medical imaging. The proposed framework offers a scalable, robust solution for brain tumor diagnosis, paving the way for improved disease prediction and treatment planning in clinical practice.